Re: [RFC] new MSR r/w functions per CPU

[Dave Jones <davej@redhat.com>]

On Wed, Dec 13, 2006 at 02:19:52PM -0800, H. Peter Anvin wrote:

 > > void rdmsr_on_cpu(unsigned int cpu, unsigned long msr, unsigned long *lo, unsigned long *hi)
 > > {
 > >     cpumask_t oldmask;
 > > 
 > >     oldmask = current->cpus_allowed;
 > >     set_cpus_allowed(current, cpumask_of_cpu(cpu));
 > > 
 > > 	rdmsr(msr, &lo, &hi);
 > > 
 > >     set_cpus_allowed(current, oldmask);
 > > }
 > > 
 > 
 > [The above doesn't work, by the way.  This approach was discussed a long 
 > time ago, and vetoed due to the potential for deadlock.]

Can you explain this a little further? I'm fairly certain
there are places in the kernel already doing this (or similar).
In fact, I cut-n-pasted most of the above from similar code in the
powernow-k8 driver.  What exactly can we deadlock on?

		Dave

-- 
http://www.codemonkey.org.uk